<h2>Must know</h2>
<ul>
<li><strong>Databricks launches Lakebase for AI workloads</strong>: Databricks launched Lakebase, a new PostgreSQL-compatible database specifically designed for AI workloads. It aims to merge operational database capabilities with their existing data lakehouse architecture, offering a serverless OLTP database with features like instant data branching and pgvector support. This move by Databricks, leveraging technology from Neon and Mooncake Labs, signifies a push to integrate transactional data directly into AI platforms without complex ETL processes, currently available on AWS with Azure in preview. <a href="https://app.daily.dev/posts/56RHAShlB" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google embedding model deprecation causes vendor lock-in</strong>: Google deprecated its <code>text-embedding-004</code> model, rendering existing vectors unusable for a production RAG-based movie recommender. This highlights a critical issue: different embedding models create incompatible vector spaces, making migration difficult and exposing a hidden form of vendor lock-in. The solution involved re-embedding the entire dataset and a blue/green deployment, underscoring the need to treat embedding models as infrastructure and design for change from the outset. <a href="https://app.daily.dev/posts/xhDI1zj4i" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Linux kernel 7.0 rc1 released</strong>: The Linux kernel 7.0 rc1 has been released, with Linus Torvalds humorously attributing the major version bump to his dislike of large numbers, rather than a significant feature milestone. This release solidifies Rust support, brings faster cache clearing, non-disruptive kernel updates, and updated AMD/Intel silicon support. It also includes performance improvements for RISC-V and LoongArch, alongside the removal of a legacy IBM ThinkPad modem driver. <a href="https://app.daily.dev/posts/cYP8biu57" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Uber and OpenAI unveil traffic control platforms</strong>: Uber and OpenAI revealed their infrastructure-level traffic control platforms, “Global Rate Limiter” and “Access Engine” respectively, designed to manage massive request volumes and ensure fair access across their systems. <a href="https://app.daily.dev/posts/OVoWP2oNj" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Fresh File Explorer extension for VS Code</strong>: A new VS Code extension, ‘Fresh File Explorer’, replaces the standard file explorer to display recently modified files, integrating Git history and offering features like a recency heatmap and one-click restore for deleted files. <a href="https://app.daily.dev/posts/pQedWHBRB" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>RwLock can be slower than Mutex in Rust</strong>: Rust’s <code>RwLock</code> performed significantly worse than <code>Mutex</code> in a read-heavy LRU tensor cache due to cache-line ping-pong from atomic reader counter increments, suggesting <code>Mutex</code> might be better for very short critical sections. <a href="https://app.daily.dev/posts/rFMs5918e" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Shuru: microVM sandbox for AI agents on macOS</strong>: Shuru is a new local-first microVM sandbox for macOS, designed for AI agent workloads, providing secure, ephemeral, and reproducible Linux environments using Apple’s Virtualization.framework without Docker. <a href="https://app.daily.dev/posts/TqgsBjEoV" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Real-time glint particle rendering technique</strong>: A new research technique renders millions of microscopic glint particles in real-time at over 280 FPS on consumer GPUs, generating particle positions on-the-fly to reduce memory requirements and support UV-free surface mapping. <a href="https://app.daily.dev/posts/rT496vKCA" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>ChatDetector uses LLMs for API misuse detection</strong>: ChatDetector, an LLM-powered system, detects resource management API misuse in software by employing a ReAct-inspired framework and two-dimensional prompting, identifying 115 security bugs across six popular libraries. <a href="https://app.daily.dev/posts/UGo4Ea1t8" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Ubuntu 26.04 LTS to default to OpenJDK 25</strong>: The upcoming Ubuntu 26.04 LTS will default to OpenJDK 25, aligning its LTS release cycle with OpenJDK and bringing improvements in memory efficiency, garbage collection, and new language features. <a href="https://app.daily.dev/posts/acjUVfZBa" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Anthropic’s Claude compiler video: more hype than reality?</strong>: Anthropic released a marketing video showcasing Claude autonomously building a C compiler capable of compiling Linux, SQLite, Lua, and Doom. Critics, however, are calling out the “dishonest framing,” suggesting the real goal was merely a stress test of Claude’s ability to run autonomously for an extended period, not a straightforward demonstration of compiler-building prowess. Because, you know, AI building a compiler is just so mundane, they had to spice up the narrative. <a href="https://app.daily.dev/posts/vGYk0DbTO" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Neiry’s bio-drones: pigeons with neural interfaces</strong>: Neiry, a deeptech company, is testing ‘bio-drones’—live pigeons fitted with neural interfaces, solar panels, and cameras—to guide them along preset routes for infrastructure inspection and search-and-rescue. They claim “privacy compliance” through on-device processing, which is a bold statement for pigeons with cameras strapped to their heads. Because nothing says “privacy” like a brain-controlled bird recording everything. <a href="https://app.daily.dev/posts/m7v8MA0Dh" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Meta patents AI for deceased user simulation</strong>: Meta was granted a patent for an AI system designed to simulate a user’s social media activity, including posting and chatting, using an LLM trained on their past data, specifically for when a user is ‘absent’ from the platform. Because what could possibly go wrong with an AI impersonating deceased users? Nothing creepy about that at all, just good old-fashioned digital immortality, courtesy of Meta. <a href="https://app.daily.dev/posts/AeW0pHCcd" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
